Harden agent file exports and owner reports

[ai-christianson]

Summary

• Make filespace persistence failures actionable and bounded while preserving existing content on failed overwrites; the queue-level /app/media mount is fixed separately in gobii-ai/gobii#293.
• Restore complete, readable owner reports by treating bounded entity sets as coverage contracts, modeling entities and relationships in SQLite, and using proportionate visual hierarchy without forcing report formatting into short chat or outreach.
• Replace mock-only CSV coverage with real filespace and seeded SQLite exports, align two incoherent eval matchers, and move existing exact-image-text guidance beside the create_image prompt argument where it is applied.

Root cause

• The CSV trajectory recorded 12 PermissionError failures creating /app/media on the interactive worker. The same valid 48-row CSV succeeded after handoff to the default worker, proving a queue-specific missing mount rather than malformed CSV.
• The portfolio trajectory retrieved all 18 companies, including Gobii, and mentioned Gobii during reasoning. It then omitted only Gobii from the final answer and produced a repetitive 36-bullet catalog. Retrieval worked; coverage reconciliation and output hierarchy failed.

Verification

• Local tests: 122/122 core regression tests, 20/20 SQLite/recruitment eval tests, and 31/31 image/Sheets eval tests.
• Complexity guardrails pass exactly at 246,844 source lines. Representative prompt profiles remain six bytes smaller than main (53,627 / 51,341 / 54,130 bytes).
• DeepSeek V4 Flash trajectory regressions passed: owner portfolio 12/12 (e2a63395-1350-4a58-89d5-60642cbcfa62), real CSV persistence 6/6 (8ae0eacf-45fe-4af4-90e9-753db700a022), and current-company completeness 24/24 (fdf8b43c-431d-40b5-9dfe-e3fceb45f7bb).
• Canonical DeepSeek suite c7816134-e233-4edd-880d-b55857148f17 executed all 299 scenarios: 1,444/1,450 assertions passed. All six misses were trajectory-audited: four were stochastic model variance, one exposed an overly narrow Sheets vocabulary matcher, and one exact-image-text omission reproduced and received a scoped tool-schema hardening. Every missed case subsequently passed; the full image-generation suite passed 20/20 (b155f3fe-2e28-456a-8254-1e0c0395046f), and the exact-text case remained 4/4 on final head (b1343855-8976-456f-93b8-6c3040b64e3e). No judge or call budget was weakened to hide genuine failures.
• Exact-head CI passed: 17 successful checks, two intentional skips, all ten backend shards green.
• Exact-head preview deployed through workflow 29696564042. Argo was synced/healthy; all seven deployments were ready with zero active-pod restarts; the interactive worker had /app/media; and homepage, health, login, app shell, auth redirect, and exact-release asset smoke checks passed.

Scope

The runtime diff is 25 additions and 21 deletions across five files; the rest is regression evidence. This does not redesign filespace storage or mandate visual formatting for simple chat and outreach.

[gemini-code-assist]

Code Review

This pull request refactors filespace storage persistence error handling, updates agent system instructions and output rules, and enhances evaluation scenarios. Specifically, it introduces structured, retryable error responses (filespace_storage_unavailable) when file saving fails, updates prompt guidance to emphasize proportionate structure and flexible hierarchy, simplifies input validation in the create_csv tool, and adds a new message quality evaluation scenario for portfolio updates. Unit tests and complexity budgets have been updated accordingly. There are no review comments to address.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.